Online Appendix

CMBS and Conflicts of Interest: Evidence fromOwnership Changes for Servicers

Maisy Wong∗

University of Pennsylvania

October 2016

A Empirical appendix

A.1 Data constructionDeals

The deal dataset comes from Realpoint, downloaded on November 11th 2010. The raw datasetconsists of 1082 deals, securitized between July 1991 and November 2010. The deals affiliatedwith the United States government (N=186) and deals originated in Canada or have a trustee inCanada (N=60) are dropped. I also drop deals securitized after 2008 as these deals tend to havedifferent governance structures. This results in a final sample of 787 deals.

All securitized loans

The raw data has a total of 141,976 loans coming from the same 1082 deals in the Realpoint dealdata. Each loan has an identifier for the deal it belongs to and an unique loan ID. After droppingthe deals discussed above, the full sample of securitized loans includes 120,495 loans.

Realpoint does not report historical values for most loan attributes. Each month, it reports fixedat-origination loan attributes for all loans. For attributes that change over time (delinquency status,current LTV, current DSCR, current balance), Realpoint only updates and report the updated valuefor loans with a positive balance that month.

∗Wharton Real Estate. 3620 Locust Walk, 1464 SHDH, Philadelphia, PA 19104-6302. Email:maisy@wharton.upenn.edu.

A-1

For time-varying attributes, I construct a loan-report month level panel for 71,000 loans thathave a positive loan balance in November 2010. The ownership changes span November 2009(Berkadia) to September 2010 (CW Capital). Therefore, the first relative month where I havecurrent loan information for Berkadia is month 11. Month 26 is the last relative month whereI have current loan information for CW Capital. For liquidated loans, I downloaded one cross-section of data in January 2010, which allows me to track current LTV and current DSCR for loansliquidated between January and November 2010. However, I do not have monthly data betweenJanuary and November 2010, and cannot construct a loan panel of current loans between these twomonths.

Realized loss dataset

This is the core dataset used to estimate the effect on loan loss rates. In a separate dataset calledthe realized loss report, Realpoint lists the history of all loans that have realized losses to theCMBS trust. The realized loss report includes a sample of 11,332 liquidated loans with liquidationmonths ranging from September 1997 to November 2012. I dropped around 1400 loans withmissing liquidation dates. The primary estimation sample includes 9,272 loans liquidated between2003 and 2012.

Bloomberg data on liquidated loans

In August 2016, I downloaded auxiliary data from Bloomberg that includes 12,000 loans liquidatedfrom 2000. The data reports the termination date, special servicer, and realized losses in dollars,but not the loan balance before losses.

Bloomberg data on bond-level loss rates

In April 2016, I also downloaded data from Bloomberg that reports bond-level loss rates. Bloombergreports the cumulative loss rates for each bond, up to April 2016. Bloomberg did not have lossrates for a few bonds. I supplemented this with bond-level loss rates from Realpoint (I have somebond-level loss rates from February 2011). The conclusions are similar with and without the sup-plemented data. In total, I have bond loss rates for 14,000 bonds, with 7,000 bonds with an originalrating of A or better, and 4,000 bonds with an original rating of AAA or better.

Variable construction

The loss rates at the loan level are reported as a ratio of the total realized loss for the loan (indollars) divided by the loan balance before losses. Some outliers appear to be loans from the sameportfolio where the realized loss in dollars for the entire portfolio was divided by the loan balancefor each loan, resulting in loss rates that seem implausibly high. The core analysis winsorizesloan loss rates at the top 1% to drop these outliers. I checked the other loans with high loss rates

A-2

against other sources to determine they are not data errors. As a robustness check, I also reportanother specification where I further winsorize loss rates at the top 1%. For a subset of loans inthe loan-month panel (downloaded after November 2010), I check that the special servicer doesnot change from month-to-month. For my estimation sample, I only observe switching for fewerthan 100 liquidated loans. In the analysis, the month of liquidation is centered around event dates.All ownership changes happened within the span of a year (December 2009 to September 2010).Centered months of liquidation and calendar months are not very different. Additionally, the initialDSCR and initial LTV are winsorized at the top 1%. Property age is set to missing if the year theproperty was built is before 1700 or after 2010.

A-3

A.2 Additional resultsA.2.1 Different time controls

Table A1 repeats the main specification but uses coarser time controls instead of month of liquida-tion fixed effects. Column 1 includes quarter of liquidation fixed effects (centered around the eventdates), column 2 includes year of liquidation fixed effects and column 3 includes a post indicatorand quadratic monthly time trends. The estimated effects are slightly larger but not statisticallydifferent from the main specification (8 p.p.).

Table A1: Effect on loan loss rates with different time controls

Specification: Quarter FE Year FE Time trend(1) (2) (3)

Post × Ownership change 0.10*** 0.10*** 0.10***

( 0.03) ( 0.03) ( 0.03)

N 9272 9272 9272

R2 0.12 0.11 0.11

Quarter FE Y N N

Year FE N Y N

Trend N N Y

Special servicer FE Y Y Y

Controls Y Y Y

* p<0.1, ** p<0.05, *** p<0.01

A-4

A.2.2 Heterogeneous samples

Table A2 estimates the effect of ownership changes on loan loss rates for heterogeneous samples.Given the smaller sample sizes, instead of including both special servicer and month of liquidationfixed effects, I only include special servicer fixed effects, a post indicator, and quadratic timetrends. Column 1 presents the results for the full sample, but using time trends only. Columns 2to 6 repeat this specification but restricted to loans with balloon payments, fixed interest rates, forapartments, offices, and retail properties respectively.

Table A2: Effect of ownership changes on loan loss rates for heterogeneous samples

All Balloon Fixed rate Apartment Office Retail(1) (2) (3) (4) (5) (6)

Post × Ownership change 0.10*** -0.01 0.11*** -0.07** 0.08* 0.09**

( 0.03) ( 0.04) ( 0.04) ( 0.03) ( 0.05) ( 0.04)

N 9272 6372 8073 2444 1301 1876

R2 0.11 0.13 0.11 0.21 0.13 0.17

Time trend Y Y Y Y Y Y

Special servicer FE Y Y Y Y Y Y

Controls Y Y Y Y Y Y

* p<0.1, ** p<0.05, *** p<0.01

A-5

A.2.3 Trends in pre-determined loan attributes for liquidated loans

Each figure repeats the main loan-level analysis (equation 1) with annual estimates of the differ-ences between treated and placebo servicers (β ), controlling for month and servicer fixed effects(without loan controls). Standard errors are double clustered by servicer and month and 95%confidence intervals are included. The omitted group is year -6.

Importantly, the trend for the balloon indicator (second panel in Figure A-1) is decreasing,which indicates that fewer balloon loans are liquidated over time. While the cross-sectional anal-ysis indicates treated loans are more likely to have hotel, office, or retail loans, the trends do notindicate that more of these loans are being liquidated in the post period. Relative to year -6, thetrend for initial loan balance is also not increasing steadily over time. Overall, the trends arerelatively stable and not large enough to explain the 8 p.p. effect after ownership changes.

Figure A-1: Trends in fixed rate loans, balloon loans, and year of securitization

-0.100.000.100.200.300.40

1(Fi

xed

rate

)

-5 -4 -3 -2 -1 0 1 2

-0.20

-0.10

0.00

0.10

0.20

1(B

allo

on)

-5 -4 -3 -2 -1 0 1 2

-3-2-10123

Sec

uriti

zatio

n ye

ar

-5 -4 -3 -2 -1 0 1 2

A-6

Figure A-2: Trends in hotel, apartment, and retail loans

-.2

-.1

0

.1

1(H

otel

)

-5 -4 -3 -2 -1 0 1 2

-.2-.1

0.1.2

1(A

partm

ent)

-5 -4 -3 -2 -1 0 1 2

-.2-.1

0.1.2.3

1(R

etai

l)

-5 -4 -3 -2 -1 0 1 2

A-7

Figure A-3: Trends in industrial loans, office loans, and loan balance

-0.10

-0.05

0.00

0.05

1(In

dust

rial)

-3 -2 -1 0 1 2

-0.10-0.050.000.050.10

1(O

ffice

)

-3 -2 -1 0 1 2

-4-20246

Loan

bal

ance

(mill

ions

)

-3 -2 -1 0 1 2

Notes: The omitted group pools together years -4 to -6, to reduce the number of coefficients to estimate doubleclustered standard errors.

A-8

Figure A-4: Trends in number of properties and property age

-3

-2

-1

0

1

Num

ber o

f pro

perti

es

-5 -4 -3 -2 -1 0 1 2

-20

-10

0

10

20

Pro

perty

age

-5 -4 -3 -2 -1 0 1 2

A-9

Figure A-5: Trends in initial LTV and initial DSCR

-15

-10

-5

0

5

Initi

al L

TV

-5 -4 -3 -2 -1 0 1 2

-.2

-.1

0

.1

.2

Initi

al D

SC

R

-5 -4 -3 -2 -1 0 1 2

A-10

A.2.4 Exposure to housing bust markets

Table A3 tests whether treated loans are more likely to be located in housing bust markets. I firstcalculate rate of change in prices between 2006 and 2009, using the Federal Housing FinanceAgency (FHFA) house price index (Federal Housing Finance Agency, 2016). Column 1 definesbust markets as MSA’s with below-median growth rates, column 2 uses the 25th percentile, andcolumn 3 includes indicators for sand states (MSA’s in Nevada, Florida, Arizona, and Inland MSA’sin California). Panel A includes all securitized loans where I could merge with the FHFA data.Panel B includes liquidated loans only. Some loans have multiple MSA’s (these are dropped).Reassuringly, treated loans are not significantly more exposed to housing bust markets.

Table A3: Exposure to bust markets

Dependent variable: Bust (p50) Bust (p25) Sand States

(1) (2) (3)

Panel A: All loans

Ownership change -0.005 -0.02 -0.001

( 0.02) ( 0.02) ( 0.01)

N 101514 101514 101514

R2 0.00002 0.0003 0.000003

Panel B: Liquidated loans

Ownership change -0.01 0.01 -0.01

( 0.03) ( 0.04) ( 0.02)

N 8065 8065 8065

R2 0.00003 0.00003 0.00009

* p<0.1, ** p<0.05, *** p<0.01

A-11

A.2.5 60-day Delinquency

Column 1 of Table A4 in the appendix shows that treated loans are 0.2% more likely to become60-day delinquent, relative to a mean of 0.3% and a standard deviation of 1%. The dependentvariable is the share of servicer i’s loans (in dollars) in month t that first become 60-day delinquentin month t. This is a servicer-month level analysis with month fixed effects and robust standarderrors. The higher 60-day delinquency rate for treated servicers is a potential concern, though itwould have been ideal to test if this difference is greater relative to the pre-period.

Importantly, columns 2 to 5 in Table A4 provide loan-level analyses demonstrating that the 60-day delinquency is unlikely to be the main driver of the higher loan loss rates. Column 2 regressesan indicator that is 1 if the loan ever becomes 60-day delinquent in my sample period (EverDelinq)on the treatment dummy, controlling for all loan controls used in my main specification. Thesample includes all loans that are current in November 2010 (when I first obtained access to thedata) and standard errors are clustered at the special servicer level. The coefficient on the treatmentdummy (0.005) is insignificant and small relative to the mean (0.06). Column 3 shows that theresults are similar if I weight observations by their current balances.

Moreover, the last two columns show that loans that become 60-day delinquent in my sampledo not have higher loss rates. Column 4 regresses the loss rate on EverDelinq (including controlsin my main specification) and finds that these 60-day delinquent loans have loss rates that are 19p.p. lower. Column 5 shows that the triple interaction (post, treat, EverDelinq) is not positive andsignificant.

Therefore, while the servicer-month level analysis indicates treated servicers are more likelyto have loans that become 60-day delinquent, the loan-level analysis suggests this difference isunlikely to explain the higher loan loss rates I find above.

A-12

Table A4: DelinquencyDependent variable: Fraction Ever Ever Loss Loss

becoming delinquent delinquent rate rate

delinquent

Unit: Servicer-month level Loan level Loan level Loan level Loan level

(1) (2) (3) (4) (5)

Ownership change 0.002*** 0.005 0.01

( 0.0006) ( 0.01) ( 0.01)

Ever delinquent -0.19*** -0.18***

( 0.02) ( 0.03)

Post × Ownership change 0.10**

( 0.04)

Post × Change × Ever delinquent 0.004

( 0.02)

N 495 73545 73545 5169 9272

R2 0.03 0.01 0.01 0.13 0.15

Mean of Dependent Var. 0.003 0.06 0.06 0.51 0.50

Month FE Y N N Y Y

Servicer FE Y N N Y Y

Controls N Y Y Y Y

* p<0.1, ** p<0.05, *** p<0.01

A-13

A.2.6 Trends in outcomes

The top panel repeats the main loan-level analysis (equation 1) with annual estimates of the differ-ences between treated and placebo servicers (β ), controlling for month fixed effects, servicer fixedeffects, and loan controls. Standard errors are double clustered by month and by servicer. Theomitted group is year -6.

The bottom panel reports annual estimates from the servicer-month level analysis from Bloomberg(equation 2), for the effect of ownership changes on the monthly volume of losses. The controlsinclude month and servicer fixed effects, with robust standard errors. The omitted group is year -6and the Bloomberg data has a longer post period.

Figure A-6: Trend in effect on loan loss rates and volume of losses

-.4

-.2

0

.2

.4

Loss

rate

-5 -4 -3 -2 -1 0 1 2Relative years

-2-1

01

2Ln

(Los

ses)

-5 -4 -3 -2 -1 0 1 2 3 4 5Relative years

A-14

A.2.7 Assessing importance of selection on unobservables

I follow Altonji, Elder, and Taber (2005) and Oster (2016) to calculate how large selection on un-observables will need to be to explain away the entire 8 p.p. effect on loan loss rates. This methoddepends on the stability of coefficients (β ) and on how useful the controls are at explaining the vari-ation in the outcome (R2). I compare an uncontrolled specification (regressing loss rates on post,treat, and the interaction, β ) to the controlled specification (my most saturated specification in col-umn 3 of Table 2 in the paper). Following Oster (2016), I calculate that δ = [ βC

(βU−βC)]∗ ( RC−RU

(0.3×RC)),

where U denotes the uncontrolled specification and C denotes the controlled specification. In-tuitively, δ is high if the coefficients are stable (denominator of the first term in the bracket issmall), or the R-squared changes a lot (numerator in the second term proxies for how useful thecontrols are). In my context, β is more than halved (0.19 to 0.08) because the market conditionsalso changed significantly around 2010. Importantly, the R-squared increases more than seven-fold(0.05 to 0.38), which suggests the controls are useful.

I calculate that δ is 2.1, twice as large as the heuristic cutoff of 1. This implies that selectionon unobservables needs to be twice as important as selection on observables to explain the entire 8p.p. effect.

Table A5: Assessing importance of selection on unobservables

Specification (U)ncontrolled (C)ontrolled

(1) (2)

β 0.19 0.08

R2 0.05 0.38

N 9272 9272

δ 2.1

Month FE N Y

Special servicer FE N Y

Controls N Y

Post × Controls N Y

MSA-year FE N Y

A-15

A.2.8 Dropping years right before and after ownership changes (Bloomberg)

Table A6 demonstrates that the larger losses for loans liquidated by treated servicers after theownership changes remain using a longer post period. This analysis uses data downloaded fromBloomberg for loans liquidated between 2000 and August 2016. The data from Bloomberg onlyreports loan losses (the numerator for loan loss rates). I estimate equation 2 at the servicer-monthlevel, with robust standard errors.

Column 1 includes the full Bloomberg sample (6 years pre and post), column 2 restricts to myprimary sample period (6 pre years, 3 post years), columns 3 to 5 respectively drop 1 to 3 yearsaround the ownership changes (from the full sample in column 1). Reassuringly, the estimates arenot statistically different from each other.

The coefficients in the Bloomberg analysis (servicer-month level) are not directly comparableto the loss rate analysis (loan level). Here, Lossit is greater when the loss rate is greater (Table 2)or when the volume of liquidation is greater (column 3 in Table 6) .

Table A6: Effect on volume of losses (longer post period from Bloomberg)

Dependent variable: Ln(Lossit)

Years: 6 pre, 6 pre, Drop Drop 2 Drop 3

6 post 3 post 1 year years years

(1) (2) (3) (4) (5)

Post × Ownership change 0.84*** 0.87*** 0.84*** 0.60* 0.92**

( 0.21) ( 0.22) ( 0.28) ( 0.34) ( 0.40)

N 1201 875 929 713 521

R2 0.52 0.56 0.51 0.49 0.45

Month FE Y Y Y Y Y

Special servicer FE Y Y Y Y Y

* p<0.1, ** p<0.05, *** p<0.01

A-16

A.2.9 Bounding exercise for stockpiling effect

Assumptions needed to bound the stockpiling effect

Below are the assumptions for step 1 of the bounding exercise.

1a: I assume the most the treated servicer can liquidate each month is $101 million (ie. the heightfor Area T corresponds to the maximum from the post trend for the treated servicer).

1b: Second, I assume that the counterfactual step function jumps up 12 months before 0 (ie. thewidth for Area T).

• The ownership changes happened between December 2009 and September 2010.

• A duration of 12 months in the pre-period is plausible. In reality, the 60-plus delin-quency rate was still trending down through 2008 and only started rising towards theend of 2008. The delinquency rate was 0.88% in December 2008 and 3.5% in Septem-ber 2009.

• Moreover, a majority of delinquent loans are not liquidated.

1c: The actual losses for the treated servicer during this time period amount to $504 million (areaunder the trend line for the treated servicer, between month -12 and month 0).

1d: So, the additional losses implied by Area T = $101m∗12−$504m = $708m

1e: For the placebo servicer, the losses corresponding to Area C = $22m∗12−$96m = $168m.

The assumed heights for Area T and Area C are likely an upper bound. For example, a height of$101 million for the treated servicer implies that losses would be 2.4 times larger than actual losses(101∗12

504 = 2.4). Similarly, the implied losses for the placebo would be 2.8 times larger (22∗1296 = 2.8).

In other words, this counterfactual assumes the servicers would have more than doubled the lossesimmediately. Instead, if we assumed the height was equal to the stabilized level after the excessmass ($85 million per month for treated and $18 million for placebo), the share explained bystockpiling would be 15% (the stockpiling effect would be ($85 million - $18 million)*12 months- ($504 million - $96 million)).

No bunching pattern with conditional trends

Figure A-7 below shows that the conditional differences in the volume of losses do not exhibita bunching pattern in the post period (controlling for month fixed effects and special servicer fixedeffects). Here, I am estimating equation 2 in the paper, but with losses as the dependent variable(instead of logs), and annual estimates for the key parameter (β ). The omitted group is year -6.

A-17

Lossity =5

∑y=−5

βyOwnershipChangei ×Postity + τt +δi + εity (1)

The lack of any spike between year 0 and year 3 indicate that the main specification likelydifferences out part of the stockpiling effect. In reality, most servicers were blindsided by thecrisis and needed time to overcome their capacity constraints.

Figure A-7: Conditional trends in monthly volume of losses

-50

050

100

Loss

es in

milli

ons

-5 -4 -3 -2 -1 0 1 2 3 4 5Relative years

A-18

B Special Servicer BackgroundThis section provides more details about special servicers in the treated and placebo groups.

Berkadia Berkadia’s origins can be traced back to 1994, when GMAC Commercial MortgageCorporation was established. In December 2009, Berkshire Hathaway and Leucadia National Cor-poration bought Capmark Financial and renamed it Berkadia Commercial Mortgage LLC. One ofBerkadia’s business lines includes "proprietary lending" to "originate loans for its own balancesheet" (Berkadia, 2016). Berkadia also has a mortgage brokerage business (Leucadia NationalCorporation, 2014).

C-III In March 2010, Island Capital acquired the commercial mortgage loan servicing businessof Centerline Holdings Company and renamed it C-III. Centerline’s origins date back to 1972when it was founded as a subsidiary of Related Companies, providing multifamily finance andinvestment management services.

Island Capital also owns C-III Investment Management LLC, which manages over $3.8 billionin assets. According to their website, C-III Investment Management funds target "commercialreal estate equity, distressed commercial real estate mortgage loans, unrated and below investmentgrade commercial mortgage-backed securities (CMBS), collateralized debt obligations (CDOs),whole loans, B-notes and mezzanine debt" (C-III Capital Partners, 2016). Island Capital also hasaffiliated brokerages (C-III Realty, NAI Global) and a titling agency (Zodiac).

LNR LNR began as an operating unit within Lennar Corporation, a national homebuilder. InJuly 2010, LNR was recapitalized by a consortium of investors including Cerberus, Vornado, Oak-tree, iStar, and Aozora. All of these firms are active investors in the commercial property market(Oaktree, 2014; Istar, 2015; Cerberus, 2015). Vornado is a large Real Estate Investment Trust, andOaktree counts distressed debt and real estate as its major asset class. The primary business seg-ments of i-Star include "real estate finance, net leasing, operating properties and land". Cerberushas over $20 billion under management invested in four primary strategies, including "[d]istressedsecurities and assets (mortgage-based securities, corporate debt, non-performing loans, structuredloans) [and] real estate-related investments". At that time, Cerberus also owned 49% of AozaraBank. In April 2013 (after my sample period), Starwood Property Trust announced it had acquiredLNR.

CW Capital CW Capital was founded in 1972 as a regional, multifamily lender, and a primaryservicer. Prior to their sale, CW Capital was owned by a Canadian pension fund manager. InSeptember 2010, Fortress Investment Group LLC acquired CW Capital. Fortress has approxi-mately $67.5 billion of assets under management as of December 2014, with some investments in"distressed and undervalued assets...including...real estate" (Fortress, 2015).

B-1

Placebo group As discussed in the main text, special servicers in the placebo group have fewerself-dealing conflicts. Going across the columns in Table A7, treated servicers are more likely to beaffiliated with brokerages and online auction platforms. Brokerage and auction fees are relativelyfixed, with firms relying more on volume to drive revenue. Some of these businesses are also newbrands. Gaining market share and becoming known as the dominant intermediary in the market isimportant as customers tend to go to firms with large networks and deal flows. Therefore, the largescale of the treated servicers complement these business lines.

By contrast, none of the placebo servicers have affiliated brokerages and online auction plat-forms. Some are part of banks which were heavily regulated after 2008 and generally do not havebrokerage services (Midland, Keycorp, Wells Fargo, Washington Mutual, and Crown Northcorp.GE Capital was systemically important). According to their rating agency reports, several placeboservicers (Situs, Torchlight, Trimont) also claim that they do not engage affiliated service providersto avoid conflicts.

Second, all servicers are lenders as this is how they got into servicing to begin with. But,treated servicers are more likely to be new lenders (Berkadia and C-III are new brands), relativeto placebo servicers which are established brands. Investing in new relationships and brand equityare relatively more important for new lenders.

Third, all treated servicers are potential buyers of CMBS liquidations. In particular, the parentsof LNR, CW Capital, and C-III have distressed debt investment funds. The parents of Berkadia(Berkshire and Leucadia) do not have distressed debt investment funds per se (to my knowledge),but they are involved in acquisitions. By contrast, only three placebo servicers are potential buyers.The rest do not have proprietary investments (they are banks or the rating agency reports indicatethey do not have affiliated buyers).

Some placebo servicers also underwent ownership changes (Wachovia and Washington Mu-tual), but the new owners do not have affiliated service providers. Dropping the liquidations bythese two servicers lead to an identical effect (8 p.p. effect on loss rates). Helios and ING Clarionwere renamed during my sample period. The result is identical after dropping these servicers. Oneof the special servicers in the placebo group, JER Partners, was acquired by Island Capital towardsthe end of my sample period (August 20, 2011). The analysis does not include liquidations by JERthat are after August 2011 ($1.7 billion). The results are similar if I drop all liquidations from JER.

B-2

Table A7: AffiliatesSpecial servicer Brokerage Online auction Titling Lender Potential buyersBerkadia Treated Yes ‐ ‐ Yes YesLNR Treated ‐ Auction.com ‐ Yes YesCWCapital Treated REDS Auction.com ‐ Yes YesC‐III Treated NAI Global, 

C‐III RealtyReal Capital Markets Zodiac Title Insurance Yes Yes

Midland Placebo ‐ ‐ ‐ Yes ‐GE Capital Placebo ‐ ‐ ‐ Yes ‐Crown Northcorp Placebo ‐ ‐ ‐ Yes ‐Helios/Situs Placebo ‐ ‐ ‐ Yes ‐Orix Placebo ‐ ‐ Yes YesING Clarion/Torchlight Placebo ‐ ‐ ‐ Yes YesKeycorp Placebo ‐ ‐ ‐ Yes ‐CNL Financial Services Placebo NA NA NA Yes YesWashington Mutual Placebo ‐ ‐ ‐ Yes ‐Wachovia/Wells Fargo Placebo ‐ ‐ ‐ Yes ‐

Notes: I could not find information on whether CNL has affiliated brokerages or auction platforms (NA indicates notavailable). CW Capital has a minority interest in Auction.com (renamed Ten-X).

ReferencesAltonji, Joseph, Todd Elder, and Christopher Taber. 2005. “Selection on Observed and Unobserved

Variables: Assessing the Effectiveness of Catholic Schools.” Journal of Political Economy113 (1):151–184.

Berkadia. 2016. “Lines of Business.” Accessed online at berkadia.com.

C-III Capital Partners. 2016. “Our Businesses.” Accessed online at c3cp.com.

Cerberus. 2015. “Cerberus Capital Management, L.P., Who We Are.” Accessed online at cerber-uscapital.com.

CWCapital. 2016. “Real Estate Services.” Accessed online at cwcapital.com.

Federal Housing Finance Agency. 2016. “House Price Index.”

Fitch Ratings. 2011. “Fitch Upgrades LNR’s Special-Servicer Rating.”

———. 2016a. “C-III Asset Management, LLC Servicer Report.”

———. 2016b. “CWCapital Asset Management LLC, Servicer Report.”

Fortress. 2015. “Fortress Overview.” Accessed online at fortress.com.

Istar. 2015. “Istar Financial, Our Business.” Accessed online at istarfinancial.com.

Leucadia National Corporation. 2014. “Annual Report 2014.”

LNR. 2016. “Real Estate Servicing.” Accessed online at transparency/lnrpartners.com.

B-3

Midland Loan Services. 2016. “Midland Loan Services.” Accessed online at pnc.com.

Oaktree. 2014. “About Oaktree.” Accessed online at oaktreecapital.com.

Orix. 2016. “CMBS and Other Securitized Investments.” Accessed online at orix.com.

Oster, Emily. 2016. “Unobservable Selection and Coefficient Stability: Theory and Evidence.”The Journal of Business Economics and Statistics Forthcoming.

Situs. 2016. “Services.” Accessed online at situs.com.

Standard and Poor’s. 2010. “Servicer Evaluation: Midland Loan Services.”

———. 2015a. “Operational Risk Assessments: Situs Asset Management LLC and Situs Hold-ings, LLC.”

———. 2015b. “Servicer Evaluation: Berkadia Commercial Mortgage LLC.”

———. 2015c. “Servicer Evaluation: Torchlight Loan Services LLC.”

———. 2016a. “Servicer Evaluation: ORIX Asset Management and Loan Services Corp.”

———. 2016b. “Servicer Evaluation: Trimont Real Estate Advisors Inc.”

Torchlight Investors. 2016. “About Us.” Accessed online at torchlightinvestors.com.

Trimont. 2016. “About Us.” Accessed online at trimontrea.com.

B-4